# The Claude Code Founder Reveals The Workflow He Learned The Hard Way

Source: https://www.youtube.com/watch?v=B-UXpneKw6M
Recap page: https://rapidrecap.app/video/B-UXpneKw6M
Generated: 2026-01-09T14:35:33.179+00:00

---
## Quick Overview

Boris Cherny, the creator of Claude Code, details his advanced workflow which emphasizes giving Claude a verification feedback loop, utilizing subagents for automation, sharing configuration files like CLAUDE.md and .mcp.json with his team, employing Plan Mode for complex tasks, and utilizing hooks to enforce code quality standards automatically.

**Key Points:**
- The most critical factor for great results with Claude Code is providing a verification feedback loop, which can 2-3x the quality of the final result (00:25).
- Cherny runs 5-10 Claudes in parallel, numbering terminal tabs 1-5 and using system notifications to track when input is needed (05:21).
- He heavily uses subagents like code-simplifier and verify-app to automate common workflows, viewing them as essential for most Pull Requests (08:50).
- For team consistency, a single CLAUDE.md file containing project instructions, structure, and conventions is checked into Git and shared across the team (02:38).
- For long-running tasks, he avoids `--dangerously-skip-permissions` and instead uses permission rules in `.claude/settings.json` to pre-allow safe commands (03:26).
- He starts complex sessions in 'Plan Mode' (shift+tab twice) to iterate on the plan with Claude before switching to 'auto-accept edits mode' for execution (05:05).
- He prefers using the larger Opus 4.5 model for thinking/planning because it is faster overall than smaller models like Sonnet due to less steering required (06:15).

![Screenshot at 00:06: Boris Cherny introduces himself as the creator of Claude Code and explains his intention to demonstrate his setup, highlighting the importance of showing his configuration.](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-00-06.jpg)

**Context:** This video features Boris Cherny, the creator of Claude Code, detailing the advanced, power-user workflow he developed for using Claude as an AI coding assistant effectively. The context revolves around maximizing productivity, ensuring code quality, and enabling team-wide consistency when integrating AI into software development processes, often through custom configurations and multi-agent orchestration.

## Detailed Analysis

Boris Cherny, the creator of Claude Code, outlines several advanced techniques for maximizing productivity and quality when using Claude for coding. The cornerstone of his success is providing Claude with a verification feedback loop (Tip 13, 00:25), which he states increases final output quality by 2-3x. This verification can range from running a simple bash command to testing an entire application. He runs 5-10 Claude sessions in parallel, managing them via numbered terminal tabs and system notifications (05:21). For automation, he relies on subagents like `code-simplifier` and `verify-app` to handle repetitive tasks common in PR workflows (08:37). Team consistency is enforced by checking a shared `CLAUDE.md` file into Git, which dictates project structure, tech stack, and development workflows (02:38). For security, he explicitly avoids `--dangerously-skip-permissions` and instead configures specific allowed commands in `.claude/settings.json` (03:26). When tackling large tasks, he initiates 'Plan Mode' (shift+tab twice) to collaboratively refine the plan with Claude before switching to 'auto-accept edits mode' for execution (05:05). He also notes a preference for the larger Opus 4.5 model for complex planning because, despite being slower initially, it requires less steering, resulting in a faster overall completion time compared to smaller models (06:15). Finally, he uses a `PostToolUse` hook to automatically run formatters like `bun run format` after code edits, ensuring code style compliance in CI (07:31).

### Core Principle

- Feedback Loop: Providing Claude a way to verify its work increases final result quality by 2-3x
- Verification methods vary by domain (bash command, test suite, simulator) (00:25)

### Workflow Automation

- Uses several subagents (code-simplifier, verify-app) to automate common PR workflows
- Commands are defined via slash commands stored in `.claude/commands/` (08:17, 08:28)

### Team Consistency

- A single, version-controlled `CLAUDE.md` file dictates project instructions, tech stack, and structure for all team members (02:38)

### Permissions Management

- Avoids `--dangerously-skip-permissions` by explicitly allowing safe bash commands in `.claude/settings.json`, which is shared with the team (03:26)

### Execution Modes

- Starts complex tasks in 'Plan Mode' for iterative refinement, then switches to 'auto-accept edits mode' for execution (05:05)

### Model Preference

- Prefers Opus 4.5 for thinking because it requires less steering, leading to faster overall task completion than smaller models like Sonnet (06:15)

### Code Quality Enforcement

- Uses a `PostToolUse` hook configured in a settings file to automatically run formatters (`bun run format`) after edits, handling the last 10% of formatting errors before CI (07:31)

![Screenshot at 00:00: The Twitter profile of Boris Cherny \(@bcherny\), creator of Claude Code, showing his username and follower count.](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-00-00.jpg)
![Screenshot at 00:06: A Twitter thread post from Boris Cherny detailing his workflow, specifically mentioning the need for a verification feedback loop \(Tip 13\).](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-00-06.jpg)
![Screenshot at 00:16: The terminal output showing Claude successfully running tests, where one test passed and another failed, illustrating the feedback loop in action.](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-00-16.jpg)
![Screenshot at 01:08: The content of Cherny's CLAUDE.md file, detailing project instructions, structure, tech stack, and development workflow commands.](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-01-08.jpg)
![Screenshot at 03:29: A demonstration of the danger of not using permission controls, showing an rm -rf ./dist command resulting in 'PRODUCTION DOWN' after deleting production data.](https://ss.rapidrecap.app/screens/B-UXpneKw6M/00-03-29.jpg)
